An examination of the impact of imputation methods on the performance of deep learning models

In this thesis study, the performance of imputation methods in filling missing data and the estimation performance of deep learning methods in time series analysis were examined using Turkish Airlines, Lufthansa Airlines, Delta Airlines stock prices and synthetic dataset based on simulation. In the study, analyzes were conducted using LSTM, GRU, RNN and Transformer models. Missing data was generated at the rate of 5%, 15% and 25% in the dataset, and these missing data were filled with linear, spline, Stineman, average and random imputation techniques. The results obtained show that the performance of each model varies depending on the imputation technique used and the rate of missing data. It has been observed that the performance of deep learning methods, in particular, varies as the rate of missing data increases and depending on the imputation technique used. Among the analysis methods, the Transformer model generally performed better than other models; Among the imputation methods, the Stineman imputation technique performed well for dataset with sudden changes and sharp turns, and the Spline imputation technique provided superior results in non-linear relationships. The performances of imputation methods and deep learning methods in time series analyzes were examined comparetively and in combination. Important findings are presented regarding which methods and pairs of methods may be suitable to improve model performances.